Binlog Servers at Booking.com
35
Binlog Servers at Booking.com Jean-François Gagné jeanfrancois DOT gagne AT booking.com Presented at Percona Live Amsterdam 2015
-
Upload
jean-francois-gagne -
Category
Technology
-
view
519 -
download
0
Transcript of Binlog Servers at Booking.com
Binlog Servers
at
Booking.com
Jean-François Gagné
jeanfrancois DOT gagne AT booking.com
Presented at Percona Live Amsterdam 2015
Booking.com
1
Booking.com‟
● Based in Amsterdam since 1996
● Online Hotel and Accommodation Agent: ● 170 offices worldwide
● +801.000 properties in 221 countries
● 42 languages (website and customer service)
● Part of the Priceline Group
● And we use MySQL: ● Thousands (1000s) of servers, ~85% replicating
● >110 masters: ~25 >50 slaves & ~8 >100 slaves
2
Binlog Server: Session Summary
1. Replication and the Binlog Server
2. Extreme Read Scaling
3. Remote Site Replication (and Disaster Recovery)
4. Easy High Availability
5. Other Use-Cases
(Crash Safety and Parallel Replication)
6. Binlog Servers at Booking.com
7. New master without touching slaves
3
Binlog Server: Replication
● One master / one or more slaves
● The master records all writes in a journal: the binary logs
● Each slave:
● Downloads the journal and saves it locally (IO thread): relay logs
● Executes the relay logs on the local database (SQL thread)
● Could produce binary logs to be itself a master (log-slave-updates)
● Replication is:
● Asynchronous lag
● Single threaded (in MySQL 5.6) slower than the master
4
Binlog Server: Booking.com‟‟
● Typical replication deployment:
+---+ | M | +---+ | +------+-- ... --+---------------+-------- ... | | | | +---+ +---+ +---+ +---+ | S1| | S2| | Sn| | M1| +---+ +---+ +---+ +---+ | +-- ... --+ | | +---+ +---+ | T1| | Tm| +---+ +---+
● Si and Tj are for read scaling
● Mi are the DR master
5
Intermediate Master
:-(
Binlog Server: What
● Binlog Server (BLS): is a daemon that:
● Downloads the binary logs from the master
● Saves them identically as on the master
● Serves them to slaves
● A or X are the same for B and C:
● By design, the binary logs served by A and X are the same
6
+---+ / \
| A | ---> / X \
+---+ -----
| |
+---+ +---+
| B | | C |
+---+ +---+
Binlog Server: Read Scaling
● Typical replication topology for read scaling:
+---+
| M |
+---+
|
+--------+--------+--- ... ---+
| | | |
+---+ +---+ +---+ +---+
| S1| | S2| | S3| | Sn|
+---+ +---+ +---+ +---+
● When there are too many slaves, the network of M is overloaded:
● 100 slaves x 1Mbit/s: very close to 1Gbit/s
● OSC or purging data in RBR becomes hard
Slave lag or unreachable master for writes
7
Binlog Server: Read Scaling‟
● Typical solution: fan-out with Intermediary Masters (IM):
+---+
| M |
+---+
|
+--------+- ... -+
| | |
+---+ +---+ +---+
| M1| | M2| | Mm|
+---+ +---+ +---+
| | |
+-- ... +- ... +-- ... --+
| | | |
+---+ +---+ +---+ +---+
| S1| | T1| | Z1| | Zi|
+---+ +---+ +---+ +---+
8
● But Intermediate Masters bring problems:
● log-slave-updates IM are slower than slaves
● Lag of an IM all its slaves are lagging
● Rogue transaction on IM infection of all its slave
● Failure of an IM all its slaves stop replicating
(and action must be taken fast)
Binlog Server: Read Scaling‟‟
● Solving IM problems with shared disk:
● Filers (expensive) or DRBD (doubling the number of servers)
● sync_binlog=1 + trx_commit=1 slower replication lag
● After a crash of an Intermediate Master:
● we need InnoDB recovery replication on slaves stalled lag
● and the cache is cold replication will be slow lag
● Solving IM problems with GTIDs:
● They allow slave repointing at the cost of added complexity :-|
● But they do not completely solve the lag problem :-(
● And we cannot migrate online with MySQL 5.6 :-( :-(
9
Binlog Server: Read Scaling‟‟‟
● New Solution: replace IM by Binlog Servers +---+
| M |
+---+
|
+----------------+------ ... ------+
| | |
/ \ / \ / \
/ I1\ / I2\ / Im\
----- ----- -----
| | |
+------+ ... +--- ... +--- ... ---+
| | | | |
+---+ +---+ +---+ +---+ +---+
| S1| | S2| | Si| | Sj| | Sn|
+---+ +---+ +---+ +---+ +---+
If BLS fails, repoint slaves to other BLSs (easy by design)
10
Binlog Server: Remote Site
● Typical deployment for remote site:
+---+
| A |
+---+
|
+------+------+---------------+
| | | |
+---+ +---+ +---+ +---+
| B | | C | | D | | E |
+---+ +---+ +---+ +---+
|
+------+------+
| | |
+---+ +---+ +---+
| F | | G | | H |
+---+ +---+ +---+
● E is an IM same problems as read scaling
11
Intermediate Master
:-(
Binlog Server: Remote Site‟
● Ideally, we would like this:
+---+
| A |
+---+
|
+------+------+---------------+------+------+------+
| | | | | | |
+---+ +---+ +---+ +---+ +---+ +---+ +---+
| B | | C | | D | | E | | F | | G | | H |
+---+ +---+ +---+ +---+ +---+ +---+ +---+
● No lag and no Single Point of Failure (SPOF)
● But no master on remote site for writes (easy solvable problem)
● And expensive in WAN bandwidth (harder problem to solve)
12
Binlog Server: Remote Site‟‟
● New solution: a Binlog Server on the remote site:
+---+
| A |
+---+
|
+------+------+---------------+
| | | |
+---+ +---+ +---+ / \
| B | | C | | D | / X \
+---+ +---+ +---+ -----
|
+------+------+------+
| | | |
+---+ +---+ +---+ +---+
| E | | F | | G | | H |
+---+ +---+ +---+ +---+
13
Binlog Server: Remote Site‟‟‟
● Or deploy 2 Binlog Servers to get better resilience:
+---+
| A |
+---+
|
+------+------+---------------+
| | | |
+---+ +---+ +---+ / \ / \
| B | | C | | D | / X \ ------> / Y \
+---+ +---+ +---+ ----- -----
| |
+------+ +------+
| | | |
+---+ +---+ +---+ +---+
| E | | F | | G | | H |
+---+ +---+ +---+ +---+
● If Y fails, repoint G and H to X,
● If X fails, repoint Y to A and E and F to Y
14
Binlog Server: Remote Site‟‟‟ ‟
● Interesting property: if A fails, E, F, G & H converge
+---+
| A |
+---+
|
+------+------+---------------+
| | | |
+---+ +---+ +---+ / \
| B | | C | | D | / X \
+---+ +---+ +---+ -----
|
+------+------+------+
| | | |
+---+ +---+ +---+ +---+
| E | | F | | G | | H |
+---+ +---+ +---+ +---+
● New master promotion is easy on remote site
15
Binlog Server: Remote Site‟‟‟ ‟‟
● Step by step master promotion: 1. The 1st slave that is up to date can be the new master
2. “SHOW MASTER STATUS” or “RESET MASTER”,
and “RESET SLAVE ALL” on the new master
3. Writes can be pointed to the new master
4. Once a slave is up to date, repoint it
to the new master at the position of step # 2
5. Keep delayed/lagging slaves under X until up to date
6. Once no slaves is left under X, recycle it as
a Binlog Server for the new master
16
/ \
/ X \
-----
|
+-------------+
|
+---+
| G |
+---+ +---+
| F |
+---+
|
+------+
| |
+---+ +---+
| E | | H |
+---+ +---+
Binlog Server: High Availability
● This property can be used for high availability:
+---+
| A |
+---+
|
|
/ \
/ X \
-----
|
+------+------+------+------+------+------+------+
| | | | | | | |
+---+ +---+ +---+ +---+ +---+ +---+ +---+ +---+
| B | | C | | D | | E | | F | | G | | H | | I |
+---+ +---+ +---+ +---+ +---+ +---+ +---+ +---+
17
Binlog Server: Other Use-Cases
● Better Crash-Safe Replication
● http://blog.booking.com/
better_crash_safe_replication_for_mysql.html
● And Better Parallel Replication
● http://blog.booking.com/
better_parallel_replication_for_mysql.html
18
● Reminder: typical deployment at Booking.com:
+---+
| M |
+---+
|
+--- ... ---+ ... +-------------+--- ...
| | | |
+---+ +---+ +---+ +---+
| S1| | Sj| | Sn| | M1|
+---+ +---+ +---+ +---+
|
+-- ... --+
| |
+---+ +---+
| T1| | To|
+---+ +---+
19
● We are deploying Binlog Server Clusters to offload masters:
+---+
| M |
+---+
|
+-----+-----+ ... +-------------+--- ...
| | | | |
/ \ / \ +---+ +---+ +---+
/ A1\ / A2\ | Sj| | Sn| | M1|
----- ----- +---+ +---+ +---+
| | |
+-+-----+-+ ... -+-----+
| | | |
+---+ +---+ / \ / \
| S1| ... | Si| / Z1\ / Z2\
+---+ +---+ ----- -----
20
● We have in production:
● >40 Binlog Servers
● >20 BLS Clusters
● >650 slaves replicating
from Binlog Servers
● What is a Binlog Server Cluster ? ● At least 2 Binlog Servers
● Replicating from the same master
● With independent failure mode (not same switch/rack/…)
● With a Service DNS entry resolving to all IP addresses
Failure of a BLS transparent to slaves ● Thanks to DNS, the slaves connected to a failing
Binlog Server reconnect to the others
Easy maintenance/upgrade of a Binlog Server
21
|
+-----+
| |
/ \ / \
/ A1\ / A2\
----- -----
| |
+-+-----+-+
| |
+---+ +---+
| S1| ... | Si|
+---+ +---+
MariaDB MaxScale Binlog Router
● High Performance Parallel Replication with Bin Log Router: […] provides efficient replication from master to slave without the need for intermediate masters.
This plugin receives binlog records from a master and makes them available immediately to
slaves […]. You can use MaxScale‟s binlog server for replication to slaves across one or more
datacenters and deliver a high performance user experience.
22
● We are deploying BLS side-by-side with IM to reduce delay:
+---+
| M |
+---+
|
+-----+-----+ ... +-------------+-----+----------- ...
| | | | | |
/ \ / \ +---+ +---+ +---+ / \-->/ \
/ A1\ / A2\ | Sj| | Sn| | M1| / B1\ / B2\
----- ----- +---+ +---+ +---+ ----- -----
| | | | |
+-+-----+-+ ... +-+-----+-+
| | | |
+---+ +---+ +---+ +---+
| S1| ... | Si| | Tk| ... | To|
+---+ +---+ +---+ +---+
23
[email protected]‟‟‟ ‟
● We are deploying a new Data Center without IM:
+---+
| M |
+---+
|
+-----+-----+ ... +-------------+-----+----------- ... -----+
| | | | | | |
/ \ / \ +---+ +---+ +---+ / \-->/ \ / \-->/ \
/ A1\ / A2\ | Sj| | Sn| | M1| / B1\ / B2\ / C1\ / C2\
----- ----- +---+ +---+ +---+ ----- ----- ----- -----
| | | | | | |
+-+-----+-+ ... +-+-----+-+ +-+-----+-+
| | | | | |
+---+ +---+ +---+ +---+ +---+ +---+
| S1| ... | Si| | Tk| ... | To| | U1| ... | Up|
+---+ +---+ +---+ +---+ +---+ -----
24
HA with Binlog Servers
● Distributed Binlog Serving Service (DBSS):
+---+ | M | +---+ | +----+----------------------------------------------------+ | | +----+-------+-----------+-------+-----------+-------+----+ | | | | | | +---+ +---+ +---+ +---+ +---+ +---+ | S1|...| Sn| | T1|...| Tm| | U1|...| Uo| +---+ +---+ +---+ +---+ +---+ +---+
● Properties: ● A single Binlog Server failure does not disrupt the service (resilience)
● Minimise inter Data Center bandwidth requirements
● Allows to promote a new master without touching any slave
25
HA with Binlog Servers‟
● Zoom in DBSS:
| +----|---------------------------------------------------------+ | | |
| +----------------------+---------------------+ | | | | | |
| / \ / \ / \ | | / \--->/ \ / \--->/ \ / \--->/ \ | | ----- / \ ----- / \ ----- / \ | | | ----- | ----- | ----- |
+----|-------|--------------|-------|-------------|-------|----+ | | | | | |
26
HA with Binlog Servers‟‟
● Crash of the master: ● Step # 1: level the Binlog Servers (the slaves will follow)
+--------------------------------------------------------------+ | |
| | | |
| / \ / \ / \ | | / \--->/ \ / \--->/ \ / \--->/ \ | | ----- / \ ----- / \ ----- / \ | | | ----- | ----- | ----- |
+----|-------|--------------|-------|-------------|-------|----+ | | | | | |
27
<------------------- ------------------>
HA with Binlog Servers‟‟‟
● Crash of the master: ● Step # 2: promote a slave as the new master (there is a trick)
| +-------------------------------------------------|------------+ | | |
| +----------------------+---------------------+ | | | | | |
| / \ / \ / \ | | / \--->/ \ / \--->/ \ / \--->/ \ | | ----- / \ ----- / \ ----- / \ | | | ----- | ----- | ----- |
+----|-------|--------------|-------|-------------|-------|----+ | | | | | |
28
HA with Binlog Servers‟‟‟ ‟
● Crash of the master - the trick:
● Needs the same binary log filename on master and slaves
1. “FLUSH BINARY LOGS” on candidate master until its binary log filename follows the one available on the BLSs
2. On the new master: ● “PURGE BINARY LOGS TO „<latest binary log file>'”
● “RESET SLAVE ALL”
3. Point the writes to the new master
4. Make the Binlog Servers replicate from the new master
● From the point of view of the Binlog Server, the master only rebooted with a new ServerID and a new UUID.
29
New Master wo Touching Slaves
30
+\-/+ +---+
| X | | T1|
+/-\+ +---+
|
+------------------------+--------------------------------+
| |
+----+-------+-----------+-------+-----------+-------+----+
| | | | | |
+---+ +---+ +---+ +---+ +---+ +---+
| S1|...| Sn| | T2|...| Tm| | U1|...| Uo|
+---+ +---+ +---+ +---+ +---+ +---+
New Master wo Touching Slaves‟
31
● “FLUSH BINARY LOGS” in a loop is ugly (but it works)
● A “RESET MASTER at/to „binlog.00xxxx‟”
would be much nicer:
● https://bugs.mysql.com/bug.php?id=77438
● https://mariadb.atlassian.net/browse/MDEV-8355
(Already in MariaDB since 10.1.6)
Binlog Servers with Orchestrator
32
● Orchestrator is the tool we use for managing BLS
● Shlomi is talking about it later today:
● Wed. 23 Sept. 16:20 - 17:10 @ Winterthur
● https://www.percona.com/live/europe-amsterdam-
2015/sessions/managing-and-visualizing-your-replication-topologies-
orchestrator
Binlog Server: Links
● http://blog.booking.com/mysql_slave_scaling_and_more.html
● http://blog.booking.com/abstracting_binlog_servers_and_mysql_master_promotion_wo_reconfig
uring_slaves.html
● MaxScale High Performance Binlog Relay: https://mariadb.com/products/mariadb-maxscale
● HOWTO Install and Configure a MaxScale Binlog Server:
http://jfg-mysql.blogspot.com/2015/04/maxscale-binlog-server-howto-install-and-configure.html
● http://blog.booking.com/better_crash_safe_replication_for_mysql.html
● http://blog.booking.com/better_parallel_replication_for_mysql.html
● (http://blog.booking.com/evaluating_mysql_parallel_replication_2-slave_group_commit.html)
● The MaxScale mailing list (to ask questions): [email protected]
● Note: the Binlog Servers concept should work with any version of
MySQL (5.7, 5.6, 5.5 and 5.1) or MariaDB (10.1, 10.0, 5.5. 5.3, 5.2 and 5.1)
33
